Atii-Llama-3.2-3B

Atii stands for Adapter for Training Isometrically Inherently.

Model Overview

Atii-Llama-3.2-3B is a high-performance reasoning model built on top of meta-llama/Llama-3.2-3B-Instruct.

This model was trained without traditional training. Instead of running expensive multi-stage post-training or RL from scratch, it leverages the Atii method to directly transpose and adapt specialized Qwen post-training reasoning capabilities into the Llama architecture.

Key Features

• Zero-Shot Architecture Transfer: Transposes verifiable reasoning and cognitive search behaviors across model families without retraining.
• Autonomous Chain-of-Thought: Naturally initiates internal self-correction and multi-step verification inside <think> ... </think> blocks before delivering the final answer.
• Test-Time Compute Scaling: Dynamically scales its reasoning trace length based on problem difficulty (from ~800 words on elementary math to 5,000+ words on Olympiad competitions).

Verification Benchmarks (N=4 Majority Vote Consensus)

• GSM8K: 80.52% (Pass@4 Ceiling: 89.61%)
• MATH-500: 34.00% (Pass@4 Ceiling: 44.60%)
• AIME: Successfully proves and solves Olympiad competition problems over 5,000-word reasoning horizons.

Usage with vLLM

from vllm import LLM, SamplingParams

llm = LLM(
    model="BIBLIOKLEPT/Atii-Llama-3.2-3B",
    max_model_len=16384,
    dtype="bfloat16"
)

sampling_params = SamplingParams(
    temperature=0.6,
    top_p=0.95,
    max_tokens=8192,
    skip_special_tokens=False
)

prompt = "A farmer has 17 sheep, and all but 9 die. How many sheep are left?"
formatted = f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"

outputs = llm.generate([formatted], sampling_params)
print(outputs[0].outputs[0].text)

Attribution & Base Model

• Base Architecture: meta-llama/Llama-3.2-3B-Instruct
• Methodology: Atii (Adapter for Training Isometrically Inherently)